A shortcut for extracting causal graphs from sequence models does not work as claimed.
Early experiments suggested that a Mamba state-space model trained on next-step prediction would recover Granger-causal structure — the network of statistical dependencies where past values of one variable help predict another — using a simple matrix readout. A replication effort packaged the protocol into a five-stage falsification benchmark using standardized synthetic data generators, three intervention types, and real-world datasets with clear ground truth. The method-level claim failed at every stage: a plain linear bottleneck matched the Mamba approach; a tuned Lasso outperformed it on synthetic benchmarks; and on Lorenz-96, the one real dataset with unambiguous causal structure, classical methods like PCMCI and Granger ran ahead, with the bottleneck trailing.
The supposed advantage from interventional training data — confidence levels originally reported at p < 10^-5 — turned out to be roughly 60 percent a sample-size confound. The residual signal survived only under a non-standard random-forcing intervention scheme and reproduced just as well in plain bivariate Granger, meaning the effect belongs to the phenomenon, not the architecture. If classical methods already capture it, there is nothing architecture-specific to explain.
What survives is a reusable benchmark for stress-testing future causal discovery claims. The authors frame it as the paper's lasting artifact rather than the method itself — which tells you most of what you need to know about how the experiments went.